What Is Stream Processing?

Introduction to Stream Processing Trond Hagen SINTEF ICT, Applied Mathematics Geilo, January 2008 ICT 1 Schedule, Thursday 09:00 - 09:40 Introduction to stream processing Trond Hagen 09:50 - 10:30 Introduction to stream processing cont’d Trond Hagen 15:00-15-40 CUDA programming Johan Seland 15:50-16-30 CUDA programming cont’d Johan Seland 17:00-18-30 Examples of applications André Brodtkorb, Johan Seland, ICT 2 Schedule, Friday 09:00 - 09:45 Introduction to Cell BE Trond Hagen 10:00 - 10:45 Programming Cell BE André Brodtkorb 11:00-12:00 “Birds of a feather” – parallel processing Johan Seland Summary and discussion ICT 3 Thursday evening Don’t miss the quiz in the bar after the dinner! Chance to win a ~10 000,- NOK HPC graphics card sponsored by NVIDIA ICT 4 Outline Introduction to Stream Processing Introduction to Graphics Processing Units (GPUs) GPU Architecture GPU Programming Models Examples Looking Forward ICT 5 What is Stream Processing? A stream is a set of input and output data Stream processing is a series of operations (kernel functions) applied for each element in a stream Uniform streaming is most typical. One kernel at a time is applied to all elements of the stream Single Instruction Multiple Data (SIMD) ICT 6 Instruction-Based Processing Instructions Processor Data Cache Memory Memory operands (data) During processing, the data required for an instruction’s execution is loaded into the cache, if not already present. Very flexible model, but has the disadvantage that the data-sequence is completely driven by the instruction sequence, yielding inefficient performance for uniform operations on large data blocks. ICT 7 Data Stream Processing Processor Data Data Memory Memory pipeline configuration The processor is first configured by the instructions that need to be performed and in the next step a data-stream is processed. The execution is distributed among several pipelines. ICT 8 Different Computing Paradigms Data stream processing is advantageous when large data blocks undergo the same operation, because this allows the memory efficient streaming and parallel processing of the data. Example: Matrix-matrix addition C = A+B // instruction based // data stream for(i=0; i<numRows; i++) setInputArrays(A, B); for(j=0; j<numCols; j++) setOutputArrays(C); C[i][j]=A[i][j]+B[i][j]; loadKernel(”return a+b”); execute(); ICT 9 Hardware Evolution “The number of transistors on an integrated circuit for minimum component cost doubles every 24 months” - Moore’s Law Moore’s law still seems to hold true Increases in transistor density have driven roughly proportional increases in processor performance Performance gains have also relied heavily on the increase of clock frequency, which is closely related to transistor size Smaller transistors can switch faster ICT 10 Hardware Evolution (cont’d) Higher frequency gives easy and instant performance benefits Software runs faster without any modification This has allowed the computing industry to realize increasing value from their existing code base, with relatively little effort In the resent years we have not seen the traditional increase in core frequency Increasing the frequency has several implications: ICT 11 Memory Problem Memory speeds have not increased as fast as core frequencies A processor can wait through hundreds of clock cycles if it has to get data or instructions from main memory Larger caches combined with instruction level parallelism can reduce the memory-wait time ICT 12 Longer Instruction Pipelines Longer instruction pipelines allow for higher core frequencies Longer pipelines results in more cache-miss penalty and lower the number of completed instructions per cycle The Intel Pentium 4 Prescott had a pipeline length of 31 stages and a frequency > 3 GHz Industry seems to converge between 9 and 11 stages ICT 13 Power Consumption Problem When the frequency increases, the power consumption increases disproportionately The dependency between core frequency and power consumption is often said to be quadratic ICT 14 Multi-Core Processors Multi-core can be used to take advantage of the continuously increase in transistor density Doubling the number of cores and halving frequency gives roughly the same performance, while the power consumption is reduced drastically Adding cores makes it possible to get higher performance without increasing the power density ICT 15 Intel Core 2 Duo Two “fat” cores ICT 16 General-purpose CPUs ~50 % of the die area is cache An addition to that a lot of the area is used for instruction level parallelism Instruction pipelining (execution of multiple instructions can be partially overlapped) Superscalar (parallel instruction pipelines) Branch prediction (predict outcome of decisions) Out-of-order execution The amount of transistors doing direct computation is shrinking relative to the total number of transistors ICT 17 Solution: Heterogeneous Architectures? Heterogeneous computing is the strategy of using multiple types of processing elements within a single workflow, and allowing each to perform the tasks to which it is best suited. Do not need only “fat” general purpose cores. Instead one can design smaller and simpler computational units, which can very effectively perform special tasks, e.g. floating point operations. Two most important heterogeneous systems: Graphics Processing Units Cell BE ICT 18 Cell Broadband Engine One “fat” core and eight “thin” cores ICT 19 Graphics Processing Units (GPUs) ICT 20 What is a GPU? Graphics rendering device Efficient at displaying computer graphics Their highly parallel architecture makes them very effective for a range of complex algorithms Driven by demand for increased realism in games NVIDIA GeForce 8800 GTX All PCs have a GPU ICT 21 GPU-Design Premises Games typically renders 10 000s triangles @ 60 fps Screen resolution is typically 1600 x 1200 Each pixel is recalculated every frame A pixel loosely corresponds to a lightweight thread This corresponds to creating, running and destroying 115 200 000 threads per second GPUs are designed to make these operations fast ICT 22 State of the art real-time rendering ICT 23 State of the art real-time rendering ICT 23 GPU Versus CPU Performance Performance 600 ATI NVIDIA Intel 450 300 GFLOPS 150 0 37,000 37,500 38,000 38,500 39,000 ICT 24 GPU vs. CPU Performance (cont’d) Nr. 1 top 500 November 2007: BlueGene /L 212 992 processors ~596 TFLOPS Theoretically the same processing power as ~1150 high- end GPUs http://www.top500.org ICT 25 Power / Performance Ratio “If the performance per watt of today’s computers doesn’t improve, the electrical costs of running them could end up far greater than the initial hardware price tag” - google CPU – Intel Core 2 Duo “Conroe” 65 W / 20 GFLOPs = 3.25 Watt / GFLOPS GPU 150 W / 500 GFLOPs = 0.3 Watt / GFLOPS Cell BE 70 W / 250 GFLOPs = 0.28 Watt / GFLOPS (5 W per SPE) ICT 26 State of the art real-time rendering (cont’d) ICT 27 State of the art real-time rendering (cont’d) ICT 27 3D Navier-Stokes Fluid Simulator ICT 28 3D Navier-Stokes Fluid Simulator ICT 28 Deformable Skin Demo ICT 29 Deformable Skin Demo ICT 29 More Simulation and Physics in Games Human head: Sum-of-Gaussian algorithm for realistic real- time images of human skin Gaussian convolution filter for blurring, image processing. Smoke: Fluid simulator for smoke More realistic simulations in future games Frog: Real-time mesh deformation in combination with rendering is not possible on a CPU. Real-time simulations require heavy computational power ICT 30 GPU Architecture ICT 31 Overall System Architecture CPU 1333 MHz 12.0 GB/s FSB 12.0 GB/s 4 GB/s System North Bridge GPU To DRAM PCIe Display 86 GB/s, GeForce 8800 GTX South Bridge Graphics DRAM Other Peripherals ICT 32 NVIDIA GeForce 8800 GTX Price: 3000 NOK. ~520 GFLOPS performance Core clock: 675 MHz 681 million transistors Internal bandwith: 86.4 GB/s 128 stream processors 1.35 GHz Memory GDDR3 768 MB 1.8 GHz Programmable in high-level language ICT 33 GeForce 8800 Architecture Overview Image courtesy of rage3d.com ICT 34 GeForce 8800 Architecture Overview (cont’d) 16 stream processors in each multiprocessor 128 stream processors in total. L1 cache shared between all stream processors in a multiprocessor Image courtesy of rage3d.com ICT 35 GPU Programming Models ICT 36 Traditional Problems with GPU Programming The user must be an expert in computer graphics and its computational idioms to make effective use of the hardware The programming model is unusual Debugging is difficult Few libraries ICT 37 Graphics Processing Pipeline This is the abstraction the user is exposed to by the DirectX and OpenGL graphics APIs Hard to understand for non-graphics programmers ICT 38 Compute Unified Driver Architecture (CUDA) Software interface for utilizing the GPU as a data-parallel computing device developed by NVIDIA. No need for graphics API Available for NVIDIA GeForce 8000-series GPUs and the HPC Tesla-series. ICT 39 CUDA – Software Layers CPU Application CUDA Libraries CUDA Runtime CUDA Driver GPU ICT 40 CUDA – Programming Model The GPU is viewed as a device that operates as a coprocessor to the CPU (host). A kernel is a function that is executed on the device as many different threads, doing the same type of operations. The batch of threads that executes a kernel is organized as a grid of thread blocks. A thread block is a batch of threads that can communicate with each other. All threads in a block runs on the same

Load more